Why is julia-debug producing redundant debug intrinsics?

Internals & Design Internals
,
1

When the following function’s IR is generated by julia-debug,

function add_mat(a,b)
         ni,nj = size(a)
         for i=1:ni
           for j=1:nj
             a[i,j] = a[i,j] + b[i,j]
           end
         end
       end

Some of the @llvm.dbg.values are repeated, sometimes even in the next line.

[...]
 12   call void @llvm.dbg.value(metadata i8** null, i64 0, metadata !15, metadata !23), !dbg !22
 13   call void @llvm.dbg.value(metadata i8** null, i64 0, metadata !16, metadata !23), !dbg !22
[...]
 68   call void @llvm.dbg.value(metadata i64 %"#temp#1.010", i64 0, metadata !19, metadata !23), !dbg !22
 69   call void @llvm.dbg.value(metadata i64 %"#temp#.09", i64 0, metadata !17, metadata !23), !dbg !22
 70   %24 = add i64 %"#temp#1.010", -1, !dbg !29
 71   %25 = add i64 %"#temp#.09", -1, !dbg !29
 72   %26 = mul i64 %25, %6, !dbg !29
 73   %27 = add i64 %24, %26, !dbg !29
 74   %28 = getelementptr double, double* %3, i64 %27, !dbg !29
 75   %29 = load double, double* %28, align 8, !dbg !29, !tbaa !30
 76   call void @llvm.dbg.value(metadata i64 %"#temp#1.010", i64 0, metadata !19, metadata !23), !dbg !22
 77   call void @llvm.dbg.value(metadata i64 %"#temp#.09", i64 0, metadata !17, metadata !23), !dbg !22
 78   %30 = mul i64 %25, %11, !dbg !29
 79   %31 = add i64 %24, %30, !dbg !29
 80   %32 = getelementptr double, double* %8, i64 %31, !dbg !29
 81   %33 = load double, double* %32, align 8, !dbg !29, !tbaa !30
 82   %34 = fadd double %29, %33, !dbg !29
 83   call void @llvm.dbg.value(metadata i64 %"#temp#1.010", i64 0, metadata !19, metadata !23), !dbg !22
 84   call void @llvm.dbg.value(metadata i64 %"#temp#.09", i64 0, metadata !17, metadata !23), !dbg !22
[...]

add_mat.ll

The intrinsics are present even around instructions that don’t use the register the intrinsic is about.

So, why is julia-debug producing redundant debug intrinsics ?

I’m having an issue with code generating NVPTX kernels in Polly whenever it has to deal with debug information. Just guessing this might be a possible reason.

2

Which ones are repeated? They appears to be all different.

3

I’m sorry. I’ve applied the same text-formatting on repetitive metadata